diff --git a/tgt-vvp/eval_expr.c b/tgt-vvp/eval_expr.c
index a0ca10771..9458c33aa 100644
--- a/tgt-vvp/eval_expr.c
+++ b/tgt-vvp/eval_expr.c
@@ -17,7 +17,7 @@
  *    Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
  */
 #ifdef HAVE_CVS_IDENT
-#ident "$Id: eval_expr.c,v 1.72 2002/08/28 17:15:35 steve Exp $"
+#ident "$Id: eval_expr.c,v 1.73 2002/08/28 18:38:07 steve Exp $"
 #endif
 
 # include  "vvp_priv.h"
@@ -715,6 +715,27 @@ static struct vector_info draw_add_immediate(ivl_expr_t le,
       return lv;
 }
 
+/*
+ * The subi is restricted to imm operands that are <= 16 bits wide.
+ */
+static struct vector_info draw_sub_immediate(ivl_expr_t le,
+					     ivl_expr_t re,
+					     unsigned wid)
+{
+      struct vector_info lv;
+      unsigned long imm;
+
+      lv = draw_eval_expr_wid(le, wid);
+      assert(lv.wid == wid);
+
+      imm = get_number_immediate(re);
+      assert( (imm & ~0xffff) == 0 );
+
+      fprintf(vvp_out, "    %%subi %u, %lu, %u;\n", lv.base, imm, wid);
+
+      return lv;
+}
+
 static struct vector_info draw_mul_immediate(ivl_expr_t le,
 					     ivl_expr_t re,
 					     unsigned wid)
@@ -752,6 +773,16 @@ static struct vector_info draw_binary_expr_arith(ivl_expr_t exp, unsigned wid)
 	  && number_is_immediate(re, 8*sizeof(unsigned long)))
 	    return draw_add_immediate(le, re, wid);
 
+      if ((ivl_expr_opcode(exp) == '-')
+	  && (ivl_expr_type(re) == IVL_EX_ULONG))
+	    return draw_sub_immediate(le, re, wid);
+
+      if ((ivl_expr_opcode(exp) == '-')
+	  && (ivl_expr_type(re) == IVL_EX_NUMBER)
+	  && (! number_is_unknown(re))
+	  && number_is_immediate(re, 16))
+	    return draw_sub_immediate(le, re, wid);
+
       if ((ivl_expr_opcode(exp) == '*')
 	  && (ivl_expr_type(re) == IVL_EX_NUMBER)
 	  && (! number_is_unknown(re))
@@ -1760,6 +1791,9 @@ struct vector_info draw_eval_expr(ivl_expr_t exp)
 
 /*
  * $Log: eval_expr.c,v $
+ * Revision 1.73  2002/08/28 18:38:07  steve
+ *  Add the %subi instruction, and use it where possible.
+ *
  * Revision 1.72  2002/08/28 17:15:35  steve
  *  Generate %load/nx for indexed load of nets.
  *
diff --git a/vvp/codes.h b/vvp/codes.h
index 9b1764834..8f0e7a74b 100644
--- a/vvp/codes.h
+++ b/vvp/codes.h
@@ -19,7 +19,7 @@
  *    Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
  */
 #ifdef HAVE_CVS_IDENT
-#ident "$Id: codes.h,v 1.48 2002/08/28 17:15:06 steve Exp $"
+#ident "$Id: codes.h,v 1.49 2002/08/28 18:38:07 steve Exp $"
 #endif
 
 
@@ -91,6 +91,7 @@ extern bool of_SET_X(vthread_t thr, vvp_code_t code);
 extern bool of_SHIFTL_I0(vthread_t thr, vvp_code_t code);
 extern bool of_SHIFTR_I0(vthread_t thr, vvp_code_t code);
 extern bool of_SUB(vthread_t thr, vvp_code_t code);
+extern bool of_SUBI(vthread_t thr, vvp_code_t code);
 extern bool of_VPI_CALL(vthread_t thr, vvp_code_t code);
 extern bool of_WAIT(vthread_t thr, vvp_code_t code);
 extern bool of_XNOR(vthread_t thr, vvp_code_t code);
@@ -154,6 +155,9 @@ extern vvp_code_t codespace_index(vvp_cpoint_t ptr);
 
 /*
  * $Log: codes.h,v $
+ * Revision 1.49  2002/08/28 18:38:07  steve
+ *  Add the %subi instruction, and use it where possible.
+ *
  * Revision 1.48  2002/08/28 17:15:06  steve
  *  Add the %load/nx opcode to index vpi nets.
  *
diff --git a/vvp/compile.cc b/vvp/compile.cc
index 38bc65f73..1286f0574 100644
--- a/vvp/compile.cc
+++ b/vvp/compile.cc
@@ -17,7 +17,7 @@
  *    Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
  */
 #ifdef HAVE_CVS_IDENT
-#ident "$Id: compile.cc,v 1.139 2002/08/28 17:15:06 steve Exp $"
+#ident "$Id: compile.cc,v 1.140 2002/08/28 18:38:07 steve Exp $"
 #endif
 
 # include  "arith.h"
@@ -136,6 +136,7 @@ const static struct opcode_table_s opcode_table[] = {
       { "%shiftl/i0", of_SHIFTL_I0, 2, {OA_BIT1,OA_NUMBER,   OA_NONE} },
       { "%shiftr/i0", of_SHIFTR_I0, 2, {OA_BIT1,OA_NUMBER,   OA_NONE} },
       { "%sub",    of_SUB,    3,  {OA_BIT1,     OA_BIT2,     OA_NUMBER} },
+      { "%subi",   of_SUBI,   3,  {OA_BIT1,     OA_BIT2,     OA_NUMBER} },
       { "%wait",   of_WAIT,   1,  {OA_FUNC_PTR, OA_NONE,     OA_NONE} },
       { "%xnor",   of_XNOR,   3,  {OA_BIT1,     OA_BIT2,     OA_NUMBER} },
       { "%xnor/r", of_XNORR,  3,  {OA_BIT1,     OA_BIT2,     OA_NUMBER} },
@@ -1448,6 +1449,9 @@ void compile_net(char*label, char*name, int msb, int lsb, bool signed_flag,
 
 /*
  * $Log: compile.cc,v $
+ * Revision 1.140  2002/08/28 18:38:07  steve
+ *  Add the %subi instruction, and use it where possible.
+ *
  * Revision 1.139  2002/08/28 17:15:06  steve
  *  Add the %load/nx opcode to index vpi nets.
  *
diff --git a/vvp/opcodes.txt b/vvp/opcodes.txt
index 89f5e10c4..99ebda6ab 100644
--- a/vvp/opcodes.txt
+++ b/vvp/opcodes.txt
@@ -1,7 +1,7 @@
 /*
  * Copyright (c) 2001 Stephen Williams (steve@icarus.com)
  *
- *  $Id: opcodes.txt,v 1.39 2002/08/28 17:15:06 steve Exp $
+ *  $Id: opcodes.txt,v 1.40 2002/08/28 18:38:07 steve Exp $
  */
 
 
@@ -461,6 +461,13 @@ operand are x, then the entire result is x.
 
 See also the %add instruction.
 
+* %subi <bit-l>, <imm>, <wid>
+
+This instruction arithmetically subtracts the immediate value from the
+left vector. The <imm> value is a 16bit unsigned value zer-extended to
+the <wid> of the left vector. The result replaces the left vector.
+
+See also the %addi instruction.
 
 * %vpi_call <name> [, ...]
 
diff --git a/vvp/vthread.cc b/vvp/vthread.cc
index 5ab9bc6c3..c526bc05a 100644
--- a/vvp/vthread.cc
+++ b/vvp/vthread.cc
@@ -17,7 +17,7 @@
  *    Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA
  */
 #ifdef HAVE_CVS_IDENT
-#ident "$Id: vthread.cc,v 1.83 2002/08/28 17:15:06 steve Exp $"
+#ident "$Id: vthread.cc,v 1.84 2002/08/28 18:38:07 steve Exp $"
 #endif
 
 # include  "vthread.h"
@@ -2128,6 +2128,58 @@ bool of_SUB(vthread_t thr, vvp_code_t cp)
       return true;
 }
 
+bool of_SUBI(vthread_t thr, vvp_code_t cp)
+{
+      assert(cp->bit_idx[0] >= 4);
+
+      unsigned word_count = (cp->number+CPU_WORD_BITS-1)/CPU_WORD_BITS;
+
+      unsigned long*lva = vector_to_array(thr, cp->bit_idx[0], cp->number);
+      unsigned long*lvb;
+      if (lva == 0)
+	    goto x_out;
+
+      lvb = new unsigned long[word_count];
+
+      lvb[0] = ~ cp->bit_idx[1];
+      for (unsigned idx = 1 ;  idx < word_count ;  idx += 1)
+	    lvb[idx] = ~0;
+
+      unsigned long carry;
+      carry = 1;
+      for (unsigned idx = 0 ;  (idx*CPU_WORD_BITS) < cp->number ;  idx += 1) {
+
+	    unsigned long tmp = lvb[idx] + carry;
+	    unsigned long sum = lva[idx] + tmp;
+	    carry = 0;
+	    if (tmp < lvb[idx])
+		  carry = 1;
+	    if (sum < tmp)
+		  carry = 1;
+	    if (sum < lva[idx])
+		  carry = 1;
+	    lva[idx] = sum;
+      }
+
+      for (unsigned idx = 0 ;  idx < cp->number ;  idx += 1) {
+	    unsigned bit = lva[idx/CPU_WORD_BITS] >> (idx % CPU_WORD_BITS);
+	    thr_put_bit(thr, cp->bit_idx[0]+idx, (bit&1) ? 1 : 0);
+      }
+
+      delete[]lva;
+      delete[]lvb;
+
+      return true;
+
+ x_out:
+      delete[]lva;
+
+      for (unsigned idx = 0 ;  idx < cp->number ;  idx += 1)
+	    thr_put_bit(thr, cp->bit_idx[0]+idx, 2);
+
+      return true;
+}
+
 bool of_VPI_CALL(vthread_t thr, vvp_code_t cp)
 {
 	// printf("thread %p: %%vpi_call\n", thr);
@@ -2267,6 +2319,9 @@ bool of_CALL_UFUNC(vthread_t thr, vvp_code_t cp)
 
 /*
  * $Log: vthread.cc,v $
+ * Revision 1.84  2002/08/28 18:38:07  steve
+ *  Add the %subi instruction, and use it where possible.
+ *
  * Revision 1.83  2002/08/28 17:15:06  steve
  *  Add the %load/nx opcode to index vpi nets.
  *